DroneIQA-VLE Multi-Task Drone Image Quality Assessment via Vision-Language Ensemble

Jul 16, 2026

DroneIQA-VLE proposes a solution for the ICME 2026 Drone-IQA Grand Challenge on target-aware image quality assessment for low-altitude UAV imagery. The framework jointly predicts global, target, and background quality scores by ensembling two complementary pipelines: SigLIP2 vision encoders with multi-task regression heads, and a LoRA-adapted Qwen3.5-9B multimodal large language model for quality score regression. Global quality is computed by arithmetic averaging of both pipelines, achieving 2nd place in the challenge.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages4

Reading Time10 min

Introduction

Model Architecture

SigLIP2 Multi-task Models

Qwen3.5-9B Multimodal LLM

DroneIQA-VLE: Multi-Task Drone Image Quality Assessment via Vision-Language Ensemble Wei Sun 1 , Weixia Zhang2 , Hongjian Zhan 1 , Mingkai Lu2 , Yixuan Gao2 , Guangtao Zhai2 1East China Normal University, 2 Shanghai Jiao Tong University arXiv :2607 .004 16v 1 [ cs .CV] 1 Jul 2026 Abstract—We present DroneIQA-VLE, our solution to the ICME 2026 Drone-IQA Grand Challenge on Target-aware Image Quality Assessment for Low-altitude UAV Images. The framework jointly predicts global, target, and background quality scores by ensembling two complementary pipelines: (1) SigLIP2 vision encoders with multi-task regression heads, and (2) a LoRA-adapted Qwen3.5-9B multimodal large language model for quality score regression. The final global quality prediction is obtained by arithmetically averaging the outputs of both pipelines. Our method achieves 2nd place in the challenge, demonstrating its effectiveness. The code is available at [https:](https:)//[github.com/sunwei925/DroneIQA-VLE](github.com/sunwei925/DroneIQA-VLE). I. INTRODUCTION Unmanned Aerial Vehicle (UAV) imagery has become increasingly prevalent in applications such as surveillance, traffic monitoring, and emergency response. However, UAV images exhibit distinct quality characteristics compared to conventional natural images due to diverse viewpoints, small target regions, complex backgrounds, and spatially nonuniform degradations. These factors make standard image quality assessment (IQA) [1] methods less suitable for UAV scenarios. Traditional full-reference metrics such as PSNR and SSIM [2] are impractical since pristine references are unavailable in real-world UAV deployments, while most no-reference IQA methods [3]–[9] focus solely on global perceptual quality without considering target-region usability and background interference. To promote research on UAV-oriented quality modeling, the Drone-IQA GC 2026 Grand Challenge [10] introducesa target-aware benchmark comprising approximately 6,000 UAV images collected from VisDrone [11] and UAVDT [12] datasets, annotated by 18 human raters along three perceptual dimensions: global quality, target quality, and background quality. The challenge requires participants to predict the global quality score of UAV images, while target and background quality annotations can serve as auxiliary supervision. Submissions are evaluated using the average of Pearson Linear Correlation Coefficient (PLCC) and Spearman Rank Correlation Coefficient (SRCC) . In this report, we present DroneIQA-VLE, our solution to this challenge. Our approach is motivated by two key observations. First, multi-task learning with auxiliary quality dimensions (target and background) provides beneficial inductive bias for the primary global quality prediction task. Second, vision-encoder-based models and multimodal large language models capture complementary quality-relevant features: the former excels at spatial quality feature extraction, while the latter provides high-level semantic understanding and perceptual reasoning. By ensembling predictions from SigLIP2 vision encoders and a LoRA-adapted Qwen3.5-9B model, our framework achieves robust and accurate quality predictions that generalize well across diverse UAV imaging conditions. II. MODEL ARCHITECTURE We investigate two complementary modeling paradigms for multi-task drone image quality assessment: pure visionencoder regression and multimodal large language model regression. A. SigLIP2 Multi-task Models We employ two SigLIP2-based visual backbones spanning different architectural scales within the Vision Transformer family. Each backbone is equipped with three independent regression heads to predict the three quality dimensions simultaneously. SigLIP2 ViT-L/16 (384 px). The first architecture builds upon the visual encoder of SigLIP2 ViT-L/16 [13], a Vision Transformer Large model with a patch size of 16 that operates at a 384 × 384 input resolution. It produces a 1024-dimensional featur

FAQs

DroneIQA-VLE要解决什么问题？

面向低空低轨UAV的目标感知图像质量评估，需要同时预测全局质量、目标区域质量和背景质量分数，并更好地适配无人机图像的非均匀退化与复杂场景。

该方法如何同时建模全局、目标和背景三个质量维度？

使用两个互补流水线进行集成：一条基于SigLIP2视觉编码器的多任务回归头直接预测三项分数，另一条使用LoRA适配的Qwen3.5-9B多模态LLM把输入图像映射到连续质量分数，再对两路全局预测取算术平均得到最终全局质量。

模型集成时的全局质量预测是如何得到的？

最终的全局质量预测通过对两条流水线输出进行算术平均获得，以实现更稳健、更准确的质量估计。

Jul 16, 2026

Upload by Lucas Martin